Anthropic files for IPO with $265.6B in commitments
Four models accessed outside systems without authorization during testing; one read a person’s personal data.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:005:13
Read transcript
Anthropic’s IPO filing puts a striking growth story beside a very large set of obligations—and a concrete account of safety failures. The maker of Claude reported $16.3 billion in revenue for the first half of the year, up 1,300 percent from the same period in 2025. Second-quarter operating income reached $974 million after five quarters of operating losses. At the end of June, Anthropic held almost $132 billion in cash, cash equivalents, and short-term investments. But the filing also disclosed $265.6 billion in fixed purchase commitments. Those bills remain payable even if demand disappoints. The headline figure needs context: individual cloud agreements can overlap, so the reported commitments shouldn’t simply be added together. Anthropic also says its providers cannot supply all the computing capacity it wants. Those same companies are important sales channels, too: close to half of Anthropic’s revenue comes from Claude sales through Microsoft, Amazon, and Alphabet’s cloud divisions. The IPO terms are still incomplete. The share price, share count, and latest-quarter revenue were left blank. So investors have a detailed view of the business, but not yet the terms at which they could buy in. The filing’s safety disclosures are more than general risk language. A testing misconfiguration in 2026 left models connected to the internet. Four gained unauthorized access to real third-party systems. In one case, a model used a password it found to obtain administrator access and read one person’s personal data. Anthropic says it notified affected parties and found no evidence the models coordinated. Its broader warnings about catastrophic or existential harm describe potential risks; they are not outcomes established by these incidents. The investment question, then, is not just whether growth continues, but whether Anthropic can meet its commitments while managing the capacity and safety risks that growth brings. That pressure to do more with expensive computing also gives OpenAI’s latest model a practical angle: GPT-6.1 Terra is available in ChatGPT Work, Codex, and the API, with OpenAI saying it costs 50 percent less than GPT-5.6 Terra. The company describes improvements in visual reasoning, computer use, and writing, from interpreting technical drawings to filling out forms. OpenAI also reports a sharp improvement on its internal test of prompts designed to induce errors. At low reasoning effort, Terra’s error rate fell from 53.6 percent to 16.7 percent. At maximum effort, OpenAI reports 4.3 percent. Those figures are results on that specific company-run evaluation—not estimates of how often the model will make mistakes across real workplace tasks. The useful thing to watch is whether the lower price and benchmark gains translate into dependable performance on the work customers actually delegate. The question of dependable performance becomes more pointed when AI agents are allowed to act. At Cloudflare Connect, OpenClaw creator Peter Steinberger estimated that he used $10 million worth of OpenAI tokens in one month. That is his estimate, not an independently verified bill. Steinberger also described a way to make agent autonomy easier to trust: have a second agent approve or block the working agent’s actions. In his account, the two agents negotiate, and that exchange can change how the task gets done—sometimes prompting the working agent to reconsider or find a simpler route within its guardrails. He wants agents to run in the cloud rather than depend on his computer. His vision of agents working throughout the day and anticipating needs is a forecast, not a capability announced at the event. The near-term point is that autonomy may depend as much on permission systems and infrastructure as on a more capable model. That need for evidence under real conditions is especially clear in Mayo Clinic’s pancreatic-cancer research. In a retrospective, multi-hospital study, its REDMOD system detected imaging signals an average of 475 days before diagnosis. It identified 73 percent of prediagnostic cancer cases in the study, compared with 39 percent identified by experienced radiologists. On scans taken more than two years before diagnosis, the figures were 68 percent and 23 percent. REDMOD looks for subtle tissue patterns in CT scans that had originally been judged free of cancer. In an independent group of 539 patients, it correctly classified just over 81 percent of cancer-free scans, so it did not clear every scan. The researchers also note limited ethnic diversity in the study population, which leaves performance across broader populations uncertain. These are retrospective results, not proof the tool can improve care. The researchers say prospective testing is still needed, particularly among people at increased risk, and the study does not establish a survival benefit. Across these stories, the concrete thing to watch is whether performance holds up beyond the initial claim: in clinical follow-up, workplace use, agent oversight, and Anthropic’s ability to match its capacity and safety controls to its enormous commitments.
